The overall aim of this proposal is to continue to advance the accuracy of homology modeling for structural genomics. This will be done by advances in two main directions: (a) better classification of protein domains as they are discovered by experimentalists and (b) better methods to refine protein models to bring them closer to the actual structure. Classification of known structures as they are discovered will enable experimentalists to better evaluate their progress and relate it to the work of other groups. It will also provide a valuable data-base of accurate multiple structure alignments for use in both homology modeling and fold recognition. Better refinement will involve the development of energy functions that favor native protein structures at a detailed all-atom level as well as more powerful methods for moving from an initial model towards the real structure. Besides leading to more accurate homology models, both these advances will have far-reaching applications in all studies of molecular function including modeling of ligand binding, modeling of protein-protein interactions and more general simulation of protein function. We believe that such significant progress will be best achieved by successful completion of the following specific aims: (1) Develop a reliable method for domain classification based on our improved structure alignment method Structal, which performed surprisingly well in this study of six different methods. Data will be updated every week using the computer resources that we have available for this grant. (2) Use the regular Structal searches done as part of Aim (1) to produce multiple structural alignments that will be based on our methods for multiple structure superposition and alignment conflict resolution. (3) Use orthogonal normal modes in torsion angle and Cartesian space to generate thousands of near-native decoys. These decoy sets will be generated to have small local deformations but differ from the native structure by different extents of the global root mean square (RMS) deviation. (4) Use near-native decoys to test different molecular mechanics and knowledge-based energy functions. Successes and failures to discriminate structures closer to the native structure from those that are further will be used to systematically improve energy functions that can be used to refine near native structures.